World Labs Atlas is a multimodal world model from Fei-Fei Li's World Labs that turns a handful of photos into navigable 3D scenes and camera-controlled video, built for filmmakers, game designers, and robotics teams who need consistent spatial output rather than random clips.
Core Features
- Pixel-perfect camera control: feed Atlas a camera trajectory as a native geometric input, and it generates new views from any angle instead of guessing from a text prompt.
- Shared spatial context: a multimodal autoregressive diffusion transformer anchors every image to a 3D position, so generated frames stay geometrically consistent.
- Controllable long video: produce up to one minute of 1440p video from just one to six reference images with a hand-designed camera path.
- Sparse-view 3D reconstruction: rebuild real scenes from as few as two or three photos, exporting point clouds and 3D Gaussian splats.
- Space-time simulation: reframe existing video and feed Real-to-Sim pipelines that train robots from synthetic first-person data.
- Benchmarked in-house against MiniMax H3, Gemini Omni Flash, and Seedance 2.5, with external raters preferring its camera-controlled output on a blind eval.
Use Cases
- Filmmakers and VFX artists staging bullet-time and complex camera moves from a single still.
- Game and virtual-world studios generating consistent, navigable environments.
- Robotics teams building synthetic training scenes via Real-to-Sim without physical capture.
- Architects and designers visualizing unbuilt spaces from sparse on-site photos.
Pricing
Atlas is in early access for a small group of unnamed partners; World Labs has not published pricing or a general-availability date. It will power future versions of the company's Marble product.
Our Take
Best for studios that already live in 3D pipelines and want director-grade camera control from a few photos; the trade-off is that there is no self-serve API yet, so you cannot build on Atlas unless you are in the partner program. See other AI design tools.
